A 10 MW data center does not automatically buy 10 MW of electricity every hour. It may use less than its installed capacity, and whatever the IT equipment does consume has to be supported by cooling, power conversion, pumps, fans, lighting and other facility systems.
That is why electricity cost is best understood as a chain rather than a single multiplication:
The formula is simple. Choosing honest inputs is the difficult part.
A 10 MW example shows how quickly the bill becomes material
Suppose a facility carries a steady 10 MW IT load for a full year. At a PUE of 1.50, total facility demand averages 15 MW. That works out to roughly 131.4 million kWh per year.
At an effective electricity cost of $0.08/kWh, that energy costs about $10.5 million per year. At $0.12/kWh, it becomes roughly $15.8 million. Nothing about the servers changed; only the electricity price did.
Illustrative calculation: 10 MW continuous IT load, PUE 1.50, 8,760 hours. These are modeling inputs, not quoted utility rates.
There is no single “U.S. data center electricity rate”
The latest published EIA data makes the point clearly. Through May 2026, the average U.S. retail electricity price was 13.79 cents/kWh for commercial customers and 8.83 cents/kWh for industrial customers. In May alone, those averages were 13.54 cents and 8.71 cents, respectively. EIA publishes the current year-to-date figures in its Electric Power Monthly data.
Those numbers are useful context. They are not a plug-in tariff for a 50 MW data center.
Very large data centers may be served under utility tariffs, negotiated structures or load classes that do not resemble the average bill paid by a typical commercial customer. Demand charges, transmission, riders, taxes, power-factor provisions and other tariff components can also matter.
For a real project budget, the correct input is the applicable utility or contractual rate structure at the site — not the national commercial average.
This is one place where a tidy internet answer can cause a surprisingly large error. Using 13.79 cents instead of 8 cents in the 10 MW example above changes annual modeled cost by more than $7 million. At hyperscale size, a few cents per kWh is not a rounding issue.
PUE turns IT energy into facility energy
PUE, or power usage effectiveness, expresses total facility energy divided by IT equipment energy. A PUE of 1.50 means the facility uses 1.50 units of total energy for each unit consumed by the IT load.
The Uptime Institute's 2026 global survey reports an industry-wide annual average PUE of 1.52. Its capacity-weighted calculation is lower at 1.36, reflecting the fact that many newer, larger facilities operate more efficiently; Uptime says leading new facilities routinely report PUE of 1.3 or below. Uptime published the updated analysis in August 2026.
Total facility load for 10 MW of IT.
Close to Uptime's 2026 capacity-weighted average.
Uptime's 2026 per-facility industry average.
At 10 MW of steady IT load, moving from PUE 1.52 to 1.36 reduces facility overhead by 1.6 MW. Across a full year, that is about 14 million kWh less energy. At $0.08/kWh, the modeled difference is roughly $1.12 million annually.
That does not mean every site can economically “buy” a lower PUE. Climate, redundancy, load profile, cooling design and capital cost all matter. PUE is an efficiency metric, not an instruction to pursue the lowest possible number regardless of context.
This is where I would stop anyone from comparing two facilities using PUE alone. A lower PUE is useful, but a facility with cheaper electricity and a slightly worse PUE can still have the lower energy bill.
Electricity price and PUE interact
A 0.1 change in PUE matters more in a high-cost electricity market than in a low-cost one. The same efficiency project can therefore have very different financial payback depending on where the facility operates.
Annual cost: about $7.88M
Annual cost: about $15.77M
Annual cost: about $6.83M
Annual cost: about $13.67M
In the lower-rate case, improving from PUE 1.50 to 1.30 saves about $1.05 million per year. At the higher rate, the same modeled efficiency improvement saves roughly $2.10 million.
Installed MW, contracted MW and actual load are not the same thing
Another common mistake is to take a facility described as “100 MW” and calculate 100 × 8,760 hours as if the entire capacity were continuously consumed.
Capacity describes what a site is designed or contracted to support. Actual electricity consumption depends on how much IT equipment is installed, how heavily it runs and how the facility load changes over time.
Those figures can converge in a mature, heavily utilized facility, but they can differ substantially during lease-up or phased deployment. A cost model should therefore use expected load by period rather than simply applying the full design capacity from day one.
A phased build can change the first five years dramatically
Imagine a 20 MW facility that opens with 5 MW of IT load and grows over several years. Using the full 20 MW from the opening date would overstate early electricity consumption by a wide margin.
Real load ramps are rarely that smooth, but the principle is useful. Electricity cost should follow expected utilization, not just development capacity.
Why the U.S. power question is getting larger, not smaller
The scale of data center electricity demand is changing quickly. Lawrence Berkeley National Laboratory's report to Congress estimates that U.S. data centers used about 176 TWh in 2023, roughly 4.4% of total U.S. electricity consumption. Its modeled 2028 range is approximately 325–580 TWh, equivalent to about 6.7%–12% of projected U.S. electricity consumption. The LBNL report explains that the wide range reflects uncertainty around AI hardware deployment, utilization and cooling choices.
About 4.4% of U.S. electricity use.
About 6.7%–12% of projected U.S. electricity use.
That growth does not tell us what one facility will pay. It does explain why grid connection, generation, transmission and long-term power procurement have moved so high on the data center development agenda.
What belongs in a serious electricity-cost model
Expected IT load by period. Not just the ultimate design capacity.
PUE assumption. Preferably varied by season or operating condition when the model needs that level of detail.
Applicable tariff or contract. Energy price, demand charges and relevant riders should come from the actual service structure.
Escalation. A multi-year TCO model should not freeze electricity prices unless that assumption is deliberate.
Operational scenarios. High, base and low utilization usually tell more than one point estimate.
For an early screening model, it is perfectly reasonable to use a simple effective $/kWh assumption and a representative PUE. The important thing is to label them as assumptions. Once a site is being seriously evaluated, those placeholders should be replaced with the utility tariff, expected load profile and the facility's actual design basis.
One cent per kWh is worth more than it looks
At small scale, a one-cent difference feels trivial. At data center scale it is not.
The same one-cent difference over five years, before any load growth or escalation, is roughly $6.57 million.
This is why electricity belongs beside construction cost in any serious site comparison. A market can be more expensive to build in and still produce the better long-term economics if power is materially cheaper or easier to secure. The opposite can also happen.
The electricity bill ultimately comes down to four variables: how much IT load is actually running, how much facility overhead supports it, what the applicable power structure charges, and how those three things change over time. “MW × electricity price” is a useful first line. It is not yet a data center energy model.